docs: design selectable UR policy config

docs/superpowers/specs/2026-07-23-selectable-ur-policy-config-design.md

@@ -0,0 +1,45 @@
+# Selectable UR Policy Configuration Design
+
+## Goal
+
+Allow the Gradio Space to load either the state-conditioned or no-state π₀.₅
+UR configuration without editing deployment code.
+
+## Interface
+
+Add a `Policy config` dropdown with exactly two accepted values:
+
+- `pi05_ur_demo_no_state` (default);
+- `pi05_ur_demo_state`.
+
+The no-state option is the default because the currently deployed checkpoint
+repository is named `pi0.5-nostate`. The seven state controls remain visible so
+users can switch configurations without rebuilding the page. The page explains
+that these controls do not provide discrete state conditioning in no-state mode.
+
+## Loading and Caching
+
+Pass the selected configuration from `predict_ui` to `ModelManager.get`. Validate
+it against an explicit allowlist before downloading or loading a model. Extend
+the cache key from `(model_id, checkpoint_path)` to
+`(model_id, checkpoint_path, config_name)` so switching configurations always
+rebuilds the OpenPI Policy.
+
+The default loader calls `openpi_config.get_config(config_name)`. Its successful
+load status reports both `config_name` and `config.model.discrete_state_input`.
+Repository names and checkpoint paths never determine the configuration
+implicitly.
+
+## Inference and Errors
+
+Both configurations keep the existing UR image, prompt, state, action transform,
+normalization, and output shape contracts. The OpenPI model transforms decide
+whether state is used as discrete conditioning. Unsupported configuration names
+fail before model download with an actionable validation error.
+
+## Tests
+
+CPU tests verify the dropdown default and choices, config propagation through
+the UI, allowlist rejection, cache separation by configuration, and loader use
+of the selected config. Existing inference, artifact, and Space tests remain
+unchanged.